Frontiers in Small Area Estimation Research: Application to Welfare Indicators

This paper reviews the main methods for small area estimation of welfare indicators. It begins by discussing the importance of small area estimation methods for producing reliable disaggregated estimates. It mentions the baseline papers and describes the contents of the different sections. Basic direct estimators obtained from area-specific survey data are described first, followed by simple indirect methods, which include synthetic procedures that do not account for the area effects and composite estimators obtained as a composition (or weighted average) of a synthetic and a direct estimator. The previous estimators are design-based, meaning that their properties are assessed under the sampling replication mechanism, without assuming any model to be true. The paper then turns to proper model-based estimators that assume an explicit model. These models allow obtaining optimal small area estimators when the assumed model holds. The first type of models, referred to as area-level models, use only aggregated data at the area level to fit the model. However, unit-level survey data were previously used to calculate the direct estimators, which act as response variables in the most common area-level models. The paper then switches to unit-level models, describing first the usual estimators for area means, and then moving to general area indicators. Semi-parametric, non-parametric, and machine learning procedures are described in a separate section, although many of the procedures are applicable only to area means. Based on the previous material, the paper identifies gaps or potential limitations in existing procedures from a practitioner’s perspective, which could potentially be addressed through research over the next three to five years.